Updated 4 Sep 2026: Adds Cursor's September 2 Self-Hosted Machines expansion. Cloud Agents can now run tool execution on customer-controlled personal machines, dynamically scheduled enterprise pools, Kubernetes and partner sandboxes including Vercel/Cloudflare/AWS Lambda, while Cursor retains the agent loop/inference. This materially changes the Cloud Agent execution and governance boundary.
Cursor turns cloud agents into event-driven workers — and now lets teams choose where they execute
Cursor Cloud Agents can now span greenfield builds, event-driven work and customer-controlled execution infrastructure: tool calls can run on personal machines, enterprise pools, Kubernetes or partner sandboxes while Cursor keeps the agent loop and inference in its cloud.
Cursor added event subscriptions, persistent goals and isolated subagents to Cloud Agents in August 2026.
Greenfield Cloud Agents can start without an existing third-party repository and later save work into Cursor Origin.
Self-Hosted Machines expanded September 2, 2026.
My Machines can connect personal/stateful machines; Team Pools provide shared dynamically scheduled worker capacity.
Cursor documents partner/self-hosted paths for AWS Lambda, Cloudflare, Daytona, E2B, Modal, Namespace, Vercel and Coder, plus Kubernetes.
Self-hosted workers execute file, terminal, computer-use and local MCP actions on customer-managed machines.
Cursor continues to run the agent loop, inference and planning in its cloud.
Team Pool self-hosting requires Cursor Enterprise and a service-account credential.
Customers using self-hosted execution own worker infrastructure, images, secrets, scaling and production validation.
What builders should take away
Use self-hosted execution when source code, tool calls or internal services genuinely need to stay inside a controlled network; do not describe it internally as fully local AI because inference/planning still run through Cursor.
Choose My Machines for stateful personal workflows and Team Pools when capacity, service-account authentication and centrally managed images matter.
Treat partner guides as starting points: add your own worker-image patching, secrets rotation, capacity monitoring and failure recovery before making the path production-critical.
For greenfield Origin projects, decide independently where the authoritative source repository and deployment history should live; self-hosted execution does not solve source-of-truth risk.
Set concurrency and scaling policy deliberately so event-triggered agents cannot exhaust expensive customer-managed worker capacity.
Review which local MCP servers, credentials and browser sessions a self-hosted worker can reach because moving tool execution into your network can increase the sensitivity of the agent’s reachable environment.
What changed
Cursor’s Cloud Agent system has accumulated three connected changes. August releases added PR, Slack and schedule subscriptions, persistent goals, isolated subagents, Cursor Origin code hosting and a greenfield Start from scratch flow with browser preview and Vercel publication. On September 2, Cursor expanded Self-Hosted Machines so the execution side of Cloud Agents can run on infrastructure the customer manages. My Machines connects a personal laptop, VM or stateful devbox; Team Pools let enterprises expose dynamically scalable worker capacity; Cursor documents Kubernetes and partner deployment paths including AWS Lambda, Cloudflare, Daytona, E2B, Modal, Namespace, Vercel and Coder. The worker performs file edits, terminal commands, computer-use actions and local MCP calls. Cursor still runs the agent loop, inference and planning in its own cloud. Team Pools require Cursor Enterprise, while customer-managed infrastructure adds its own compute and operational cost.
Why it matters
Cloud coding agents are not one indivisible hosted service anymore. Cursor now lets teams separate the orchestration/intelligence layer from the machine that sees source code, secrets, internal services and build artifacts. That is useful for private networks, custom hardware, iOS/Mac development and compliance boundaries, but it also changes who owns reliability: when teams bring the execution environment, they own worker images, capacity, secrets, scaling policy, lifecycle and validation. Self-hosting tool execution also does not mean the entire agent runs locally because the agent loop and inference still depend on Cursor.
Cloud Agents can already wake up and keep working from events
Cursor’s August Cloud Agent work added subscriptions for pull-request events, Slack messages and schedules, persistent `/goal` context and isolated subagents. Those changes made the agent less like a one-off editor command and more like a longer-running worker around repository and team activity.
Greenfield work can begin before an external repository exists
The August 27 Start from scratch flow lets an agent create a project without a connected third-party SCM, preview its live environment in the browser and then save successful work into an Origin repository. A connected Vercel account can publish the result.
Self-Hosted Machines moves tool execution into customer infrastructure
Cursor’s September 2 update lets Cloud Agents run their tools on machines a customer controls. Personal My Machines can attach existing laptops or VMs; enterprise Team Pools are queues of workers that can scale with demand and are not tied to a single repository. Cursor also documents Kubernetes and multiple sandbox/cloud partners as deployment targets.
The trust boundary is split, not fully local
On a self-hosted worker, repository checkout, build outputs, secrets, file edits, terminal commands, computer-use tools and local MCP servers stay on that machine. Cursor’s own documentation says the agent loop, inference and planning still run in Cursor’s cloud. Teams with a policy requiring model inference itself to remain on-prem therefore need a different architecture.
Bring-your-own execution also means bring-your-own operations
Cursor’s partner guides and templates are reference architectures. The customer owns the worker image, infrastructure, secrets, scaling policy and production validation. Cursor-managed Cloud Agents remain the recommended default for most teams; self-hosting is aimed at stronger network, hardware or compliance constraints.
Vercel is one example, not the only self-hosted runtime
Vercel’s integration runs Cursor workers inside Vercel Sandbox microVMs while Cursor supplies the harness and inference. Cursor’s own supported integration list also includes AWS Lambda, Cloudflare, Namespace, Modal, Daytona, E2B and Coder, so the broader development is an execution portability layer rather than a Vercel-exclusive feature.
Timeline
2026-08-19
Cloud Agents gain subscriptions, goals and subagents
Cursor adds event-driven triggers and longer-running agent context.
2026-08-27
Greenfield Cloud Agents can start without a repo
Agents can build first, create an Origin repository later, preview in-browser and publish through Vercel.
Whether self-hosted computer-use workflows gain finer-grained permission and credential controls.
Still unclear
Cursor says its managed Cloud Agents cover the requirements of most customers and still recommends them as the default; self-hosting adds operational complexity.
Keeping code/tool execution inside a private network does not keep prompts, planning or model inference entirely local.
Partner integrations are reference architectures and their cost, cold-start, persistence and failure characteristics vary by provider.
Origin remains an evolving code-hosting layer and its long-term governance/portability guarantees can change.
Current operational boundary: customer worker executes tools while Cursor retains agent loop/inference; includes deployment requirements and ownership.
Cursor has become a concrete example of coding-tool supplier risk: a corporate acquisition can trigger a frontier-model provider’s change-of-control rights and remove a major model family from the product even when the coding tool itself remains operational.
Pi’s first stable release is interesting less for another coding-agent version number than for what its deliberately minimal core now considers mature enough to include: MCP, code-driven tool orchestration and model routing.
The Agent Host’s environment boundary has moved from local Dev Containers to remote development hosts, making persistent coding-agent sessions more portable across real remote projects.